How I Build with AI
Claude Code · Aug 28 – Sep 27, 2026
I treat AI agents as a capable engineering team I direct. I set clear goals, split work across parallel sessions, verify results against real systems, and keep architecture, product and irreversible decisions with me.
Since August 2026
- sessions
- 60
- hours paired with Claude
- 206
- commits
- 213
- pull requests merged
- 56
Rhythm
725 prompts on the 14 days I used Claude in this snapshot.
How I work
-
14 sessions coordinating multiple agents
Orchestrate agents in parallel
I run separate agents per repo and hand plans between them, so mobile and API changes move in lockstep instead of drifting apart.
-
3,765 shell commands run by agents
Verify against real data
I don't take a result on trust. Agents run the code, replay cached routes, query logs and cross-check public data feeds, and they revise findings when the evidence disagrees.
-
28 decisions Claude handed back to me
Humans own the judgment calls
Agents surface trade-offs and pause on product, data and deployment decisions. I make those calls myself, especially anything that is hard to undo.
-
56 pull requests merged
Ship small, reviewed, green
Work lands as focused pull requests with passing tests, typecheck and lint. I push back on tangents and verbose output until the code reads like it belongs.
What the sessions look like
Kinds of sessions
- Multi-step builds 51%
- Iterative refinement 21%
- Focused tasks 19%
- Exploration 6%
- Quick questions 2%
Where agents helped most
- Multi-file changes
- Proactive fixes
- Debugging
- Explaining systems
Currently building
-
TravelDot
A mobile app that turns official road, incident and weather data for your route into a short AI travel summary, so you can decide whether the trip is worth taking before you leave.
Generated from my own Claude Code session data by a small script that runs locally and publishes only aggregate numbers. Updated September 28, 2026.